RFC2547 Convergence Characterization and Optimization

1© 2004 Cisco Systems, Inc. All rights reserved.Session NumberPresentation_ID

RFC2547 ConvergenceCharacterization and OptimizationClarence [email protected]

222© 2004 Cisco Systems, Inc. All rights reserved.Presentation_ID

RFC2547 Convergence - Requirement

• 90%: Typical requirement: <10s• 9%: More aggressive requirement: <3 to 5s

– VPN is used to transport Voice

• 1%: Very Aggressive requirement: from <1s to <50ms


What failures to consider

CE2 CE3CE1

PE1

PE2 PE3

RRA1

RRA2

RRB1

RRB2

P1 P3

P2


What mechanisms for each failure

• Core Link/Node failure: BGP inheritance of IGP Convergence

• Egress PE node failure: IGP failure discovery, IGP flooding, event-driven BGP convergence

• Egress PE-CE link failure: local link failure discovery, BGP signalling, BGP convergence

• RR failure: clusters are redundant and hence no impact on connectivity. Desire to speed up the BGP reload to minimize the duration when the cluster is non-redundant

IGP Convergence is

key

Pure BGP signalling and convergence


RFC2547 Convergence does not suffer from the counting-to-infinity problem found in the Internet

• “An Experimental Study of Internet Routing Convergence”, Craig Labovitz

– “…we show that inter-domain routers in the packet switched Internet may take several minutes to reach a consistent view of the network topology after a fault…”– “…we show that even under constrained policies, the complexity of BGP convergence is exponential with respect to the number of autonomous systems…”

• Reason: there is only one possible AS path between two customer sites. Big difference between RFC2547 and Internet use of BGP


Methodology

• Same as for the IGP Fast Convergence Project– Lead customer set requirements, design context and constraints– Black Box testing to assess behavior as seen by customer. Real traffic is used to measure the Loss of Connectivity (LoC).– White Box testing to decompose the behavior into its componentsand hence to allow for implementation optimization. IOS instrumentation is used.– UUT is in a realistic IGP/BGP setup (700 IGP nodes, 2500 IGP prefixes, 100k VPNv4 routes) and is stressed by 1Mpps and 6 BGP flaps per second– Black box and white box measurements perfectly match– 20 iterations are used for each tested scenario– Design Guide


Design Context/Constraint

• Convergence to a redundant site– loadsharing or primary/backup config

• A unique RD per PE per VPN– remote PE’s do learn the two paths, no RR hiding

• 80% of the CE’s advertise less than 200 routes• It is very rare for a CE to advertise more than 1000 routes• A typical PE selects ? route via a set next-hop

– we currently test with 2000


Core Link/Node Failure

CE2 CE3CE1

PE1

PE2 PE3

RRA1

RRA2

RRB1

RRB2

P1 P3

P2


IGP Fast Convergence sub-second is conservative

• For more details, refer to Apricot 2004 presentation– also at Nanog 29, Ripe 47

• Paper under submission


IGP Fast Convergence - Reminder

• Link Failure Detection: PoS, B2B GE, DPT; if not, BFD• Fast Origination and Flooding• SPF optimization: eg. Incremental SPF• RIB/FIB Prioritization: most important prefix first• Optimization of Download distribution and HW modify• BGP fully leverage IGP entries


Egress PE Node failure

CE2 CE3CE1

PE1

PE2 PE3

RRA1

RRA2

RRB1

RRB2

P1 P3

P2


Egress PE node failure

• Adjacent core nodes detect the failure of PE2 (Link or BFD) and flood new LSP’s advertising the failure

• PE1’s IGP converges and declares PE2 unreachable • PE1: Unreachable status of a BGP nhop triggers

BGP Convergence (ie. use PE3 instead of PE2)– “BGP Next-Hop Tracking” Feature


BGP Next-Hop Tracking

• BGP registers its next-hops with the RIB• Later, RIB notify BGP when the reachability status

of these next-hops change• Dampening algorithm is used to control how

immediate the RIB notification may trigger a BGP reaction


Blackbox MeasurementEgress PE node failure

• PE1 selects 2000 prefixes from PE2– 1000 in VPN1, 1 in VPN2, …, 1 in VPN1000

• Traffic is sent to 11 prefixes in VPN1• Sub-10s for 2000 prefixes is conservative• Sub-5s is achievable


Egress PE-CE Link failure

CE2 CE3CE1

PE1

PE2 PE3

RRA1

RRA2

RRB1

RRB2

P1 P3

P2


Egress PE-CE Link Failure

• The nhop is PE2 hence IGP + BGP NHT cannot help

• This is a “pure” BGP convergence behavior– PE2 locally detects the link failure– PE2 updates its BGP, RIB, FIB tables– PE2 sends withdraws to its RR cluster– B cluster reflects to A cluster– A cluster reflects to PE1– PE1 modifies BGP, RIB and FIB table


Egress PE-CE Link Failure - Design

• Immediate and Stable BGP reaction to Link Failure– bgp fast-external-fallover: – interface dampening

• Disable Minimum Advertisement Timer for MP-iBGP– in RFC2547 with unique RD, there is 1! Path per route. Also each VPN has different attributes hence the packing is low. Hence MAT for MP-iBGP brings no real gain.– default value of 5s would lead to a worst-case impact of 15s with two RR clustersrouter bgp

address-family vpnv4

neighbor <mp-ibgp neighbor> advertisement-interval 0


Egress PE-CE Link Failure - Design

• Optimize BGP transport goodput– Large input queue: hold-queue <1500-4000> in– Input Queue Prioriritization (automatic, 22S) (SPD)– Path MTU discovery: ip tcp path-mtu-discovery– Increase the TCP window size: ip tcp window-size – dynamic update group (automatic, 24S)– update packing optimization (automatic, 26S)


Blackbox MeasurementEgress PE-CE Link Failure

• P100(1000prefixes): 3953ms • P50(1000prefixes): 2750ms


Blackbox MeasurementEgress PE-CE Link Failure

• Data VPN: 80% of the CE’s advertise less than 200 routes. It is very rare for a CE to advertise more than 1000 routes

• Voice VPN CE’s would typically advertise < 10 prefixes

0500

100015002000250030003500400045005000

P0(P0) P50(P50) P100(P100)

112501000


RR failure within a redundant cluster

CE2 CE3CE1

PE1

PE2 PE3

RRA1

RRA2

RRB1

RRB2

P1 P3

P2


RR failure within a redundant cluster

• PE1 will discover the adj down after ~120/180s• PE1 will then switch onto the same exact path but

received from the other RR of the same cluster• No Dataplane impact • When RR comes back up, sessions must be

reestablished with all peers and clients and BGP convergence must occur

– we would like to optimize this ‘bring up’ time to minimize the non-redundancy period


RR failure within a redundant clusterDesign

• No dataplane impact– ensure that both paths are imported in the local VRF’s

• Optimization of the RR ‘bring up’– implementation optimization for BGP goodput (ie 26S)– key optimization of VPNv4 BGP table in 28S1– more CPU power means faster bring up (very cpu intensive)


RR failure within a redundant clusterMeasurement

• RR_Convergence(468750, npe400, 27S1) ~ 18 min

12.0(27)S1 NPE-400 750pfx 5clusters nopeergroups clear ip bgp *

0

20

40

60

80

100

120

140

160

0 200 400 600 800 1000 1200 1400 1600

time (sec)

neighborDownflapCountflap10CountemptyInQemptyOutQbgpTableInSynccpu5Seccpu5SecIntcpu1Mincpu5Min



• RR_Convergence(468750, npe400, 27S1) ~ 18’• RR_Convergence(468750, npeG1, 28S1) ~ 4’40’’

npe-g1 12.0(28)S Optim1 750pfx 5clusters clear ip bgp *

0

20

40

60

80

100

120

140

160

0 50 100 150 200 250 300 350 400

time (sec)

neighborDownflapCountflap10CountemptyInQemptyOutQbgpTableInSynccpu5Seccpu5SecIntcpu1Mincpu5Min



• NGE1 twice as performance as NPE400– ~ factor 2 speed up in bring up time per prefix

• Key optimization in 28S1 (lab tests show 2 to 3 factor gain)

– 468750 * 0.6ms ~ 4min40sec

time per prefix

0.0

0.2

0.4

0.6

0.8

1.0

1.2

1.4

reload clear ip bgp *tim

e (m

sec)

400-12.0(28)S_optim1 250pfx

400-12.0(28)S_optim1 500pfx

400-12.0(28)S_optim1 750pfx

400-12.0(28)S_optim11000pfx

G1-12.0(28)S_optim1 250pfx

G1-12.0(28)S_optim1 500pfx

G1-12.0(28)S_optim1 750pfx

G1-12.0(28)S_optim1 1000pfx


Conclusion

• Sub-1s for core node/link failure• No impact from RR failure

– RR bring up for 500k vpnv4 routes in 4min40’’

• PE node failure, PE-CE link failure– Prefix dependent– Sub-10s is conservative for most VPN’s– Sub-5s is achievable with careful design

• Additional ideas to further optimize…

RFC2547 Convergence Characterization and Optimization

Documents

Transcript of RFC2547 Convergence Characterization and Optimization